server : When listening on a unix domain socket don't print http:// and port (#14180)

Instead show something like this:

main: server is listening on file.sock - starting the main loop

Signed-off-by: Eric Curtin <ecurtin@redhat.com>

.devops/intel.Dockerfile

@@ -49,19 +49,23 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-RUN apt-get update \
-    && apt-get install -y \
-    git \
-    python3 \
-    python3-pip \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt \
-    && apt autoremove -y \
-    && apt clean -y \
-    && rm -rf /tmp/* /var/tmp/* \
-    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
-    && find /var/cache -type f -delete
+RUN apt-get update && \
+    apt-get install -y \
+        git \
+        python3 \
+        python3-pip \
+        python3-venv && \
+    python3 -m venv /opt/venv && \
+    . /opt/venv/bin/activate && \
+    pip install --upgrade pip setuptools wheel && \
+    pip install -r requirements.txt && \
+    apt autoremove -y && \
+    apt clean -y && \
+    rm -rf /tmp/* /var/tmp/* && \
+    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
+    find /var/cache -type f -delete
 
+ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 

CMakeLists.txt

@@ -89,6 +89,14 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
+if (NOT DEFINED LLAMA_BUILD_NUMBER)
+    set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
+endif()
+if (NOT DEFINED LLAMA_BUILD_COMMIT)
+    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
+endif()
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
@@ -155,6 +163,8 @@ if (LLAMA_USE_SYSTEM_GGML)
 endif()
 
 if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
+    set(GGML_BUILD_NUMBER ${LLAMA_BUILD_NUMBER})
+    set(GGML_BUILD_COMMIT ${LLAMA_BUILD_COMMIT})
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
@@ -204,10 +214,6 @@ endif()
 include(GNUInstallDirs)
 include(CMakePackageConfigHelpers)
 
-set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
-set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
-set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
-
 set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")

README.md

@@ -18,7 +18,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Hot topics
 
 - 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
-- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
 - A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639

common/CMakeLists.txt

@@ -23,31 +23,21 @@ if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
+        # For build-info.cpp below
+        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${GIT_DIR}/index")
     else()
         message(WARNING "Git index not found in git repository.")
-        set(GIT_INDEX "")
     endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-    set(GIT_INDEX "")
 endif()
 
-# Add a custom command to rebuild build-info.cpp when .git/index changes
-add_custom_command(
-    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
-    COMMENT "Generating build details from Git"
-    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
-            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
-            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
-    WORKING_DIRECTORY "${PROJECT_SOURCE_DIR}"
-    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
-    VERBATIM
-)
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
+configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+
 set(TARGET build_info)
-add_library(${TARGET} OBJECT build-info.cpp)
+add_library(${TARGET} OBJECT ${OUTPUT_FILE})
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

common/build-info.cpp.in

@@ -1,4 +1,4 @@
-int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/chat-parser.cpp

@@ -49,6 +49,7 @@ bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::
 
     // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
     result_.tool_calls.emplace_back(tool_call);
+
     return true;
 }
 bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
@@ -378,3 +379,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         /* .is_partial = */ found_healing_marker,
     };
 }
+
+void common_chat_msg_parser::clear_tools() {
+    result_.tool_calls.clear();
+}

common/chat-parser.h

@@ -115,4 +115,6 @@ class common_chat_msg_parser {
         const std::vector<std::vector<std::string>> & args_paths = {},
         const std::vector<std::vector<std::string>> & content_paths = {}
     );
+
+    void clear_tools();
 };

common/chat.cpp

@@ -1921,7 +1921,9 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
         if (!is_partial) {
-            throw std::runtime_error(ex.what());
+            builder.clear_tools();
+            builder.move_to(0);
+            common_chat_parse_content_only(builder);
         }
     }
     auto msg = builder.result();

common/cmake/build-info-gen-cpp.cmake

@@ -1,24 +0,0 @@
-include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
-
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
-set(OUTPUT_FILE   "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
-
-# Only write the build info if it changed
-if(EXISTS ${OUTPUT_FILE})
-    file(READ ${OUTPUT_FILE} CONTENTS)
-    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_COMMIT ${CMAKE_MATCH_1})
-    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_COMPILER ${CMAKE_MATCH_1})
-    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_TARGET ${CMAKE_MATCH_1})
-    if (
-        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
-        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
-        NOT OLD_TARGET   STREQUAL BUILD_TARGET
-    )
-        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-    endif()
-else()
-    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-endif()

convert_hf_to_gguf.py

@@ -5262,6 +5262,34 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("Dots1ForCausalLM")
+class Dots1Model(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.DOTS1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["num_experts"] = self.hparams["n_routed_experts"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+
+        if self.hparams["scoring_func"] == "noaux_tc":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        else:
+            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+        if "shared_experts" in name:
+            return [(self.map_tensor_name(name), data_torch)]
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("PLMForCausalLM")
 class PLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PLM

docs/function-calling.md

@@ -11,7 +11,7 @@ Function calling is supported for all models (see https://github.com/ggml-org/ll
   - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2
   - Functionary v3.1 / v3.2
   - Hermes 2/3, Qwen 2.5
-  - Qwen 2.5 Coder (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
+  - Qwen 2.5 Coder
   - Mistral Nemo
   - Firefunction v2
   - Command R7B

docs/multimodal.md

@@ -107,3 +107,7 @@ NOTE: some models may require large context window, for example: `-c 8192`
 (tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
 (tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
 ```
+
+## Finding more models:
+
+GGUF models on Huggingface with vision capabilities can be found here: https://huggingface.co/models?pipeline_tag=image-text-to-text&sort=trending&search=gguf

ggml/src/ggml-cuda/common.cuh

@@ -262,11 +262,11 @@ static bool cp_async_available(const int cc) {
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-    return __AMDGCN_WAVEFRONT_SIZE;
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+    return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
 [[noreturn]]

ggml/src/ggml-cuda/ssm-scan.cu

@@ -10,6 +10,8 @@ __global__ void __launch_bounds__(splitD, 2)
                  float * __restrict__ dst, const int64_t L) {
     GGML_UNUSED(src1_nb0);
     GGML_UNUSED(src2_nb0);
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     const int bidx = blockIdx.x;  // split along B
     const int bidy = blockIdx.y;  // split along D
     const int tid  = threadIdx.x;
@@ -44,16 +46,16 @@ __global__ void __launch_bounds__(splitD, 2)
     if (N == 16) {
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = A_block[(wid * warpSize + i) * stride_A + wtid];
+            float value = A_block[(wid * warp_size + i) * stride_A + wtid];
             // todo: bank conflict
             // I am always confused with how to use the swizzling method to solve
             // bank conflit. Hoping somebody can tell me.
-            smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            smem_A[(wid * warp_size + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
-            smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            float value = s0_block[(wid * warp_size + i) * stride_s0 + wtid];
+            smem_s0[(wid * warp_size + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
     }
 

ggml/src/ggml-sycl/CMakeLists.txt

@@ -142,7 +142,7 @@ else()
         FetchContent_Declare(
             ONEMATH
             GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
-            GIT_TAG c255b1b4c41e2ee3059455c1f96a965d6a62568a
+            GIT_TAG 8efe85f5aaebb37f1d8c503b7af66315feabf142
         )
         FetchContent_MakeAvailable(ONEMATH)
         # Create alias to match with find_package targets name

ggml/src/ggml-sycl/common.hpp

@@ -513,9 +513,9 @@ constexpr size_t ceil_div(const size_t m, const size_t n) {
 
 bool gpu_has_xmx(sycl::device &dev);
 
-template <int N, class T> void debug_print_array(const std::string & prefix, const T array[N]) {
+template <int N, class T> std::string debug_get_array_str(const std::string & prefix, const T array[N]) {
     if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
+        return "";
     }
     std::stringstream ss;
     ss << prefix << "=[";
@@ -526,29 +526,26 @@ template <int N, class T> void debug_print_array(const std::string & prefix, con
         ss << array[N - 1];
     }
     ss << "]";
-    GGML_SYCL_DEBUG("%s", ss.str().c_str());
+    return ss.str();
 }
 
-inline void debug_print_tensor(const std::string & prefix, const ggml_tensor * tensor,
-                               const std::string & suffix = "") {
-    if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
-    }
-    GGML_SYCL_DEBUG("%s=", prefix.c_str());
+inline std::string debug_get_tensor_str(const std::string &prefix,
+        const ggml_tensor *tensor, const std::string &suffix = "") {
+    std::stringstream ss;
+    if (LIKELY(!g_ggml_sycl_debug)) { return ss.str(); }
+    ss << prefix.c_str() << "=";
     if (tensor) {
-        GGML_SYCL_DEBUG("'%s':type=%s", tensor->name, ggml_type_name(tensor->type));
-        debug_print_array<GGML_MAX_DIMS>(";ne", tensor->ne);
-        debug_print_array<GGML_MAX_DIMS>(";nb", tensor->nb);
-        if (!ggml_is_contiguous(tensor)) {
-            GGML_SYCL_DEBUG(";strided");
-        }
-        if (ggml_is_permuted(tensor)) {
-            GGML_SYCL_DEBUG(";permuted");
-        }
+        ss << "'" << tensor->name << "':type=" << ggml_type_name(tensor->type);
+        ss << debug_get_array_str<GGML_MAX_DIMS>(";ne", tensor->ne);
+        ss << debug_get_array_str<GGML_MAX_DIMS>(";nb", tensor->nb);
+
+        if (!ggml_is_contiguous(tensor)) { ss << ";strided"; }
+        if (ggml_is_permuted(tensor)) { ss << ";permuted"; }
     } else {
-        GGML_SYCL_DEBUG("nullptr");
+        ss << "nullptr";
     }
-    GGML_SYCL_DEBUG("%s", suffix.c_str());
+    ss << suffix;
+    return ss.str();
 }
 
 // Use scope_op_debug_print to log operations coming from running a model
@@ -564,10 +561,10 @@ struct scope_op_debug_print {
             return;
         }
         GGML_SYCL_DEBUG("[SYCL][OP] call %s%s:", func.data(), func_suffix.data());
-        debug_print_tensor(" dst", dst);
+        GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" dst", dst).c_str());
         if (dst) {
             for (std::size_t i = 0; i < num_src; ++i) {
-                debug_print_tensor("\tsrc" + std::to_string(i), dst->src[i]);
+                GGML_SYCL_DEBUG("%s", debug_get_tensor_str("\tsrc" + std::to_string(i), dst->src[i]).c_str());
             }
         }
         GGML_SYCL_DEBUG("%s\n", suffix.data());

ggml/src/ggml-sycl/cpy.cpp

@@ -723,8 +723,7 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const
 
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {
     // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
-    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0,
-                                         std::string(" src0 type=") + ggml_type_name(src0->type));
+    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0, debug_get_tensor_str("\tsrc0", src0));
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -347,7 +347,7 @@ static enum ggml_status
 ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                      ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
 
     if (tensor->view_src != NULL) {
@@ -385,7 +385,7 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                 const void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
     ggml_sycl_set_device(ctx->device);
@@ -413,7 +413,7 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                                 void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
 
@@ -444,8 +444,8 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                     ggml_tensor *dst) try {
     bool is_cpy_supported = ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context;
@@ -525,7 +525,7 @@ catch (sycl::exception const &exc) {
 static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value,
                                                    size_t offset, size_t size) {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu value=%u\n", size, offset, value);
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
     SYCL_CHECK(ggml_sycl_set_device(ctx->device));
@@ -805,7 +805,7 @@ static enum ggml_status
 ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                            ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
 
     ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
@@ -891,7 +891,7 @@ ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                           ggml_tensor *tensor, const void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -947,7 +947,7 @@ ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                           const ggml_tensor *tensor, void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -3863,7 +3863,7 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                const void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3884,7 +3884,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3907,8 +3907,8 @@ static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
     bool is_cpy_supported                = dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) &&
                             ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         /*

gguf-py/gguf/constants.py

@@ -343,6 +343,7 @@ class MODEL_ARCH(IntEnum):
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
     BAILINGMOE       = auto()
+    DOTS1            = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -623,6 +624,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
+    MODEL_ARCH.DOTS1:            "dots1"
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -2044,6 +2046,30 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.DOTS1: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+    ],
     # TODO
 }
 

gguf-py/gguf/tensor_mapping.py

@@ -305,7 +305,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
-            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
+            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
         ),
 
         # Feed-forward up

include/llama.h

@@ -243,14 +243,14 @@ extern "C" {
 
     typedef bool (*llama_progress_callback)(float progress, void * user_data);
 
-    // Input data for llama_decode
+    // Input data for llama_encode/llama_decode
     // A llama_batch object can contain input about one or many sequences
     // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
     //
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
-    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
+    //            (if set to NULL, the token position will be tracked automatically by llama_encode/llama_decode)
     // - seq_id : the sequence to which the respective token belongs
     //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output

requirements/requirements-compare-llama-bench.txt

@@ -1,2 +1,3 @@
 tabulate~=0.9.0
 GitPython~=3.1.43
+matplotlib~=3.10.0

scripts/compare-llama-bench.py

@@ -19,6 +19,7 @@ except ImportError as e:
     print("the following Python libraries are required: GitPython, tabulate.") # noqa: NP100
     raise e
 
+
 logger = logging.getLogger("compare-llama-bench")
 
 # All llama-bench SQL fields
@@ -122,11 +123,15 @@ help_s = (
 parser.add_argument("--check", action="store_true", help="check if all required Python libraries are installed")
 parser.add_argument("-s", "--show", help=help_s)
 parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
+parser.add_argument("--plot", help="generate a performance comparison plot and save to specified file (e.g., plot.png)")
+parser.add_argument("--plot_x", help="parameter to use as x axis for plotting (default: n_depth)", default="n_depth")
+parser.add_argument("--plot_log_scale", action="store_true", help="use log scale for x axis in plots (off by default)")
 
 known_args, unknown_args = parser.parse_known_args()
 
 logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO)
 
+
 if known_args.check:
     # Check if all required Python libraries are installed. Would have failed earlier if not.
     sys.exit(0)
@@ -499,7 +504,6 @@ else:
 
 name_compare = bench_data.get_commit_name(hexsha8_compare)
 
-
 # If the user provided columns to group the results by, use them:
 if known_args.show is not None:
     show = known_args.show.split(",")
@@ -544,6 +548,14 @@ else:
             show.remove(prop)
         except ValueError:
             pass
+
+    # Add plot_x parameter to parameters to show if it's not already present:
+    if known_args.plot:
+        for k, v in PRETTY_NAMES.items():
+            if v == known_args.plot_x and k not in show:
+                show.append(k)
+                break
+
     rows_show = bench_data.get_rows(show, hexsha8_baseline, hexsha8_compare)
 
 if not rows_show:
@@ -600,6 +612,161 @@ if "gpu_info" in show:
 headers  = [PRETTY_NAMES[p] for p in show]
 headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
 
+if known_args.plot:
+    def create_performance_plot(table_data: list[list[str]], headers: list[str], baseline_name: str, compare_name: str, output_file: str, plot_x_param: str, log_scale: bool = False):
+        try:
+            import matplotlib.pyplot as plt
+            import matplotlib
+            matplotlib.use('Agg')
+        except ImportError as e:
+            logger.error("matplotlib is required for --plot.")
+            raise e
+
+        data_headers = headers[:-4] # Exclude the last 4 columns (Test, baseline t/s, compare t/s, Speedup)
+        plot_x_index = None
+        plot_x_label = plot_x_param
+
+        if plot_x_param not in ["n_prompt", "n_gen", "n_depth"]:
+            pretty_name = PRETTY_NAMES.get(plot_x_param, plot_x_param)
+            if pretty_name in data_headers:
+                plot_x_index = data_headers.index(pretty_name)
+                plot_x_label = pretty_name
+            elif plot_x_param in data_headers:
+                plot_x_index = data_headers.index(plot_x_param)
+                plot_x_label = plot_x_param
+            else:
+                logger.error(f"Parameter '{plot_x_param}' not found in current table columns. Available columns: {', '.join(data_headers)}")
+                return
+
+        grouped_data = {}
+
+        for i, row in enumerate(table_data):
+            group_key_parts = []
+            test_name = row[-4]
+
+            base_test = ""
+            x_value = None
+
+            if plot_x_param in ["n_prompt", "n_gen", "n_depth"]:
+                for j, val in enumerate(row[:-4]):
+                    header_name = data_headers[j]
+                    if val is not None and str(val).strip():
+                        group_key_parts.append(f"{header_name}={val}")
+
+                if plot_x_param == "n_prompt" and "pp" in test_name:
+                    base_test = test_name.split("@")[0]
+                    x_value = base_test
+                elif plot_x_param == "n_gen" and "tg" in test_name:
+                    x_value = test_name.split("@")[0]
+                elif plot_x_param == "n_depth" and "@d" in test_name:
+                    base_test = test_name.split("@d")[0]
+                    x_value = int(test_name.split("@d")[1])
+                else:
+                    base_test = test_name
+
+                if base_test.strip():
+                    group_key_parts.append(f"Test={base_test}")
+            else:
+                for j, val in enumerate(row[:-4]):
+                    if j != plot_x_index:
+                        header_name = data_headers[j]
+                        if val is not None and str(val).strip():
+                            group_key_parts.append(f"{header_name}={val}")
+                    else:
+                        x_value = val
+
+                group_key_parts.append(f"Test={test_name}")
+
+            group_key = tuple(group_key_parts)
+
+            if group_key not in grouped_data:
+                grouped_data[group_key] = []
+
+            grouped_data[group_key].append({
+                'x_value': x_value,
+                'baseline': float(row[-3]),
+                'compare': float(row[-2]),
+                'speedup': float(row[-1])
+            })
+
+        if not grouped_data:
+            logger.error("No data available for plotting")
+            return
+
+        def make_axes(num_groups, max_cols=2, base_size=(8, 4)):
+            from math import ceil
+            cols = 1 if num_groups == 1 else min(max_cols, num_groups)
+            rows = ceil(num_groups / cols)
+
+            # Scale figure size by grid dimensions
+            w, h = base_size
+            fig, ax_arr = plt.subplots(rows, cols,
+                                       figsize=(w * cols, h * rows),
+                                       squeeze=False)
+
+            axes = ax_arr.flatten()[:num_groups]
+            return fig, axes
+
+        num_groups = len(grouped_data)
+        fig, axes = make_axes(num_groups)
+
+        plot_idx = 0
+
+        for group_key, points in grouped_data.items():
+            if plot_idx >= len(axes):
+                break
+            ax = axes[plot_idx]
+
+            try:
+                points_sorted = sorted(points, key=lambda p: float(p['x_value']) if p['x_value'] is not None else 0)
+                x_values = [float(p['x_value']) if p['x_value'] is not None else 0 for p in points_sorted]
+            except ValueError:
+                points_sorted = sorted(points, key=lambda p: group_key)
+                x_values = [p['x_value'] for p in points_sorted]
+
+            baseline_vals = [p['baseline'] for p in points_sorted]
+            compare_vals = [p['compare'] for p in points_sorted]
+
+            ax.plot(x_values, baseline_vals, 'o-', color='skyblue',
+                    label=f'{baseline_name}', linewidth=2, markersize=6)
+            ax.plot(x_values, compare_vals, 's--', color='lightcoral', alpha=0.8,
+                    label=f'{compare_name}', linewidth=2, markersize=6)
+
+            if log_scale:
+                ax.set_xscale('log', base=2)
+                unique_x = sorted(set(x_values))
+                ax.set_xticks(unique_x)
+                ax.set_xticklabels([str(int(x)) for x in unique_x])
+
+            title_parts = []
+            for part in group_key:
+                if '=' in part:
+                    key, value = part.split('=', 1)
+                    title_parts.append(f"{key}: {value}")
+
+            title = ', '.join(title_parts) if title_parts else "Performance comparison"
+
+            ax.set_xlabel(plot_x_label, fontsize=12, fontweight='bold')
+            ax.set_ylabel('Tokens per second (t/s)', fontsize=12, fontweight='bold')
+            ax.set_title(title, fontsize=12, fontweight='bold')
+            ax.legend(loc='best', fontsize=10)
+            ax.grid(True, alpha=0.3)
+
+            plot_idx += 1
+
+        for i in range(plot_idx, len(axes)):
+            axes[i].set_visible(False)
+
+        fig.suptitle(f'Performance comparison: {compare_name} vs. {baseline_name}',
+                     fontsize=14, fontweight='bold')
+        fig.subplots_adjust(top=1)
+
+        plt.tight_layout()
+        plt.savefig(output_file, dpi=300, bbox_inches='tight')
+        plt.close()
+
+    create_performance_plot(table, headers, name_baseline, name_compare, known_args.plot, known_args.plot_x, known_args.plot_log_scale)
+
 print(tabulate( # noqa: NP100
     table,
     headers=headers,

src/llama-arch.cpp

@@ -72,6 +72,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
+    { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -1555,6 +1556,34 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_DOTS1,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        }
+    },
     {
         LLM_ARCH_UNKNOWN,
         {

src/llama-arch.h

@@ -76,6 +76,7 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_DOTS1,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-batch.cpp

@@ -1,8 +1,14 @@
 #include "llama-batch.h"
 
+#include "llama-impl.h"
+#include "llama-cparams.h"
+#include "llama-vocab.h"
+#include "llama-memory.h"
+
 #include <cassert>
 #include <cstring>
 #include <algorithm>
+#include <sstream>
 
 llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
     // clear empty sequences
@@ -279,17 +285,55 @@ llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple
             );
 }
 
-llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0) {
-    batch = in_batch;
-    GGML_ASSERT(batch.n_tokens > 0);
-    if (!batch.pos) {
-        assert(p0 >= 0);
-        pos.resize(batch.n_tokens);
-        for (int32_t i = 0; i < batch.n_tokens; i++) {
-            pos[i] = p0 + i;
-        }
-        batch.pos = pos.data();
+llama_batch_allocr::llama_batch_allocr() {
+    const char * LLAMA_BATCH_DEBUG = getenv("LLAMA_BATCH_DEBUG");
+    debug = LLAMA_BATCH_DEBUG ? atoi(LLAMA_BATCH_DEBUG) : 0;
+
+    seq_pos.resize(LLAMA_MAX_SEQ);
+    seq_cpl.resize(LLAMA_MAX_SEQ);
+    for (auto & cur : seq_cpl) {
+        cur.resize(LLAMA_MAX_SEQ);
     }
+}
+
+bool llama_batch_allocr::init(
+        const llama_batch & batch_inp,
+        const llama_vocab & vocab,
+        const llama_memory_i * memory) {
+    clear();
+
+    batch = batch_inp;
+
+    GGML_ASSERT(batch.n_tokens > 0);
+
+    //
+    // validate input batch
+    //
+
+    if (batch.token) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= vocab.n_tokens()) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
+                return false;
+            }
+        }
+    }
+
+    if (batch.seq_id) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= LLAMA_MAX_SEQ)) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], LLAMA_MAX_SEQ);
+                    return false;
+                }
+            }
+        }
+    }
+
+    //
+    // auto-generate missing fields
+    //
+
     if (!batch.n_seq_id) {
         n_seq_id.resize(batch.n_tokens);
         for (int32_t i = 0; i < batch.n_tokens; i++) {
@@ -297,6 +341,7 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.n_seq_id = n_seq_id.data();
     }
+
     if (!batch.seq_id) {
         seq_id.resize(batch.n_tokens + 1);
         seq_id[batch.n_tokens] = NULL;
@@ -305,12 +350,201 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.seq_id = seq_id.data();
     }
+
+    if (!batch.pos) {
+        pos.resize(batch.n_tokens);
+
+        // initialize the starting position for each sequence based on the positions in the memory
+        llama_pos p0[LLAMA_MAX_SEQ];
+        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (!memory) {
+                p0[s] = 0;
+            } else {
+                p0[s] = memory->seq_pos_max(s) + 1;
+            }
+        }
+
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+
+            pos[i] = p0[seq_id];
+
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                p0[batch.seq_id[i][s]] = pos[i] + 1;
+            }
+        }
+
+        batch.pos = pos.data();
+    }
+
     if (!batch.logits) {
         // by default return the output only for the last token
         output.resize(batch.n_tokens);
         output[output.size() - 1] = true;
         batch.logits = output.data();
     }
+
+    //
+    // compute stats
+    //
+
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        n_outputs += batch.logits[i] != 0;
+    }
+
+    // determine coupled sequences
+    // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+            seq_pos[batch.seq_id[i][s]].insert(batch.pos[i]);
+
+            if (s > 0) {
+                const llama_seq_id s0 = batch.seq_id[i][0];
+                const llama_seq_id s1 = batch.seq_id[i][s];
+
+                // mark that sequence s1 is coupled to s0
+                seq_cpl[s1][s0] = true;
+
+                // note: the other way around is not necessary for now
+                //seq_cpl[s0][s1] = true;
+            }
+        }
+    }
+
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__);
+        LLAMA_LOG_DEBUG("%s:   n_tokens  = %d\n", __func__,          batch.n_tokens);
+        LLAMA_LOG_DEBUG("%s:   token     = %p\n", __func__, (void *) batch.token);
+        LLAMA_LOG_DEBUG("%s:   embd      = %p\n", __func__, (void *) batch.embd);
+        LLAMA_LOG_DEBUG("%s:   pos       = %p\n", __func__, (void *) batch.pos);
+        LLAMA_LOG_DEBUG("%s:   n_seq_id  = %p\n", __func__, (void *) batch.n_seq_id);
+        LLAMA_LOG_DEBUG("%s:   seq_id    = %p\n", __func__, (void *) batch.seq_id);
+        LLAMA_LOG_DEBUG("%s:   logits    = %p\n", __func__, (void *) batch.logits);
+        LLAMA_LOG_DEBUG("%s:   n_outputs = %d\n", __func__, n_outputs);
+
+        if (debug > 1) {
+            int seq_id_max = 0;
+            for (int32_t i = 0; i < batch.n_tokens; ++i) {
+                for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                    for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                        seq_id_max = std::max(seq_id_max, batch.seq_id[i][s]);
+                    }
+                }
+            }
+            ++seq_id_max;
+
+            LLAMA_LOG_DEBUG("%s:   token     = [\n", __func__);
+            for (int32_t i = 0; i < batch.n_tokens; ++i) {
+                std::vector<int8_t> seq_id(seq_id_max);
+
+                for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                    seq_id[batch.seq_id[i][s]] = 1;
+                }
+
+                std::stringstream ss;
+                for (int s = 0; s < seq_id_max; ++s) {
+                    if (seq_id[s]) {
+                        ss << s%10;
+                    } else {
+                        ss << ".";
+                    }
+                }
+
+                LLAMA_LOG_DEBUG("%s:  %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n",
+                        __func__, i, batch.token[i], vocab.token_to_piece(batch.token[i]).c_str(),
+                        batch.pos[i], batch.n_seq_id[i], ss.str().c_str(), batch.logits[i]);
+            }
+            LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+
+            LLAMA_LOG_DEBUG("%s:   seq       = [\n", __func__);
+            for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) {
+                if (seq_pos[s0].empty()) {
+                    continue;
+                }
+
+                std::stringstream ss;
+                for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) {
+                    if (seq_cpl[s0][s1]) {
+                        ss << s1 << " ";
+                    }
+                }
+
+                LLAMA_LOG_DEBUG("%s:  %4d: pos = [%4d, %4d], cpl = %s\n",
+                        __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str());
+            }
+            LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+        }
+    }
+
+    //
+    // consistency checks
+    //
+
+    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_pos[s].empty()) {
+            continue;
+        }
+
+        if (memory && seq_pos_min(s) != memory->seq_pos_max(s) + 1) {
+            LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s);
+            return false;
+        }
+
+        if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
+            LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
+            return false;
+        }
+    }
+
+    if (memory) {
+        for (int32_t s0 = 0; s0 < LLAMA_MAX_SEQ; ++s0) {
+            for (int32_t s1 = 0; s1 < LLAMA_MAX_SEQ; ++s1) {
+                if (seq_cpl[s0][s1]) {
+                    if (memory->seq_pos_min(s0) != memory->seq_pos_min(s1) ||
+                        memory->seq_pos_max(s0) != memory->seq_pos_max(s1)) {
+                        LLAMA_LOG_ERROR("%s: sequence %d is coupled to %d in the input batch, but have divereged\n", __func__, s0, s1);
+                        return false;
+                    }
+                }
+            }
+        }
+    }
+
+    return true;
+}
+
+const llama_batch & llama_batch_allocr::get_batch() const {
+    return batch;
+}
+
+uint32_t llama_batch_allocr::get_n_outputs() const {
+    return n_outputs;
+}
+
+llama_pos llama_batch_allocr::seq_pos_min(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].begin();
+}
+
+llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].rbegin();
+}
+
+void llama_batch_allocr::clear() {
+    n_outputs = 0;
+
+    batch = {};
+    pos.clear();
+    n_seq_id.clear();
+    seq_id.clear();
+    output.clear();
+
+    for (auto & cur : seq_pos) {
+        cur.clear();
+    }
+
+    for (auto & cur : seq_cpl) {
+        std::fill(cur.begin(), cur.end(), false);
+    }
 }
 
 //

src/llama-batch.h

@@ -4,6 +4,7 @@
 
 #include <array>
 #include <vector>
+#include <set>
 
 // very similar to llama_batch,
 // but has more metadata about sequences
@@ -18,8 +19,8 @@ struct llama_ubatch {
     llama_token  *  token;    // [n_tokens]
     float        *  embd;     // [n_embd, n_tokens]
     llama_pos    *  pos;      // [n_tokens]
-    int32_t      *  n_seq_id; // [n_seqs] // TODO: remove, should belong to only 1 sequence
-    llama_seq_id ** seq_id;   // [n_seqs] // TODO: become llama_seq_id * seq_id;
+    int32_t      *  n_seq_id; // [n_seqs]
+    llama_seq_id ** seq_id;   // [n_seqs]
     int8_t       *  output;   // [n_tokens]
 };
 
@@ -77,16 +78,41 @@ struct llama_sbatch {
     llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false);
 };
 
-// temporary allocate memory for the input batch if needed
-struct llama_batch_allocr {
-    struct llama_batch batch;
+// a helper for sanitizing and fulfilling a batch
+class llama_batch_allocr {
+public:
+    llama_batch_allocr();
+
+    // sanitize and auto-gen missing data in the input batch
+    // memory is optional. if provided will be used to check for sequence continuity and to determine the positions
+    bool init(
+            const llama_batch & batch_inp,
+            const llama_vocab & vocab,
+            const llama_memory_i * memory);
+
+    const llama_batch & get_batch() const;
+
+    uint32_t get_n_outputs() const;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const;
+
+private:
+    void clear();
+
+    llama_batch batch;
+
+    uint32_t n_outputs;
 
     std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
+
     std::vector<llama_pos>      pos;
     std::vector<int32_t>        n_seq_id;
     std::vector<llama_seq_id *> seq_id;
     std::vector<int8_t>         output;
 
-    // optionally fulfill the batch returned by llama_batch_get_one
-    llama_batch_allocr(struct llama_batch in_batch, llama_pos p0);
+    std::vector<std::set<llama_pos>> seq_pos; // seq_pos[s]: the set of positions in sequence s
+    std::vector<std::vector<bool>>   seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
+
+    int debug;
 };

src/llama-chat.cpp

@@ -183,6 +183,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_BAILING;
     } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
+    } else if (tmpl_contains("<|endofuserprompt|>")) {
+        return LLM_CHAT_TEMPLATE_DOTS1;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -643,6 +645,21 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_DOTS1) {
+        // dots.llm1.inst (DOTS1)
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|system|>" << message->content << "<|endofsystem|>";
+            } else if (role == "user") {
+                ss << "<|userprompt|>" << message->content << "<|endofuserprompt|>";
+            } else {
+                ss << "<|response|>" << message->content << "<|endofresponse|>";
+            }
+        }
+        if (add_ass) {
+            ss << "<|response|>";
+        }
     } else {
         // template not supported
         return -1;

src/llama-chat.h

@@ -43,6 +43,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_SMOLVLM,
+    LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

src/llama-context.cpp

@@ -1,6 +1,7 @@
 #include "llama-context.h"
 
 #include "llama-impl.h"
+#include "llama-batch.h"
 #include "llama-io.h"
 #include "llama-memory.h"
 #include "llama-mmap.h"
@@ -18,7 +19,8 @@
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
-    model(model) {
+    model(model),
+    batch_allocr(std::make_unique<llama_batch_allocr>()) {
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -27,8 +29,8 @@ llama_context::llama_context(
     const auto & hparams = model.hparams;
 
     cparams.n_seq_max = std::max(1u, params.n_seq_max);
-    if (cparams.n_seq_max > LLAMA_MAX_PARALLEL_SEQUENCES) {
-        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_PARALLEL_SEQUENCES));
+    if (cparams.n_seq_max > LLAMA_MAX_SEQ) {
+        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_SEQ));
     }
 
     cparams.n_threads        = params.n_threads;
@@ -494,7 +496,7 @@ float * llama_context::get_logits() {
 }
 
 float * llama_context::get_logits_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (logits == nullptr) {
@@ -517,7 +519,7 @@ float * llama_context::get_logits_ith(int32_t i) {
         }
         if (j >= n_outputs) {
             // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return logits + j*model.vocab.n_tokens();
@@ -536,7 +538,7 @@ float * llama_context::get_embeddings() {
 }
 
 float * llama_context::get_embeddings_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (embd == nullptr) {
@@ -559,7 +561,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
         }
         if (j >= n_outputs) {
             // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return embd + j*model.hparams.n_embd;
@@ -719,40 +721,26 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch,
     return res;
 }
 
-int llama_context::encode(llama_batch & inp_batch) {
-    if (inp_batch.n_tokens == 0) {
+int llama_context::encode(const llama_batch & batch_inp) {
+    if (batch_inp.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
-    // temporary allocate memory for the input batch if needed
     // note: during encode, we always pass the full sequence starting from pos = 0
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : 0);
+    if (!batch_allocr->init(batch_inp, model.vocab, nullptr)) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+        return -1;
+    }
 
-    const llama_batch & batch = batch_allocr.batch;
-    const int32_t n_tokens = batch.n_tokens;
+    const llama_batch & batch = batch_allocr->get_batch();
 
-    const auto & hparams = model.hparams;
+    const uint32_t n_tokens = batch.n_tokens;
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    // TODO: move the validation to the llama_batch_allocr
-    if (batch.token) {
-        for (int32_t i = 0; i < n_tokens; ++i) {
-            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
-                return -1;
-            }
-
-            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
-                LLAMA_LOG_ERROR("%s: invalid seq_id[%d] = %d > %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
-                throw -1;
-            }
-        }
-    }
-
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
-    GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens");
+    GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");
 
     if (t_compute_start_us == 0) {
         t_compute_start_us = ggml_time_us();
@@ -763,6 +751,8 @@ int llama_context::encode(llama_batch & inp_batch) {
 
     n_queued_tokens += n_tokens;
 
+    const auto & hparams = model.hparams;
+
     const int64_t n_embd = hparams.n_embd;
 
     llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true);
@@ -775,7 +765,7 @@ int llama_context::encode(llama_batch & inp_batch) {
         return -2;
     };
 
-    for (int32_t i = 0; i < n_tokens; ++i) {
+    for (uint32_t i = 0; i < n_tokens; ++i) {
         output_ids[i] = i;
     }
 
@@ -831,7 +821,8 @@ int llama_context::encode(llama_batch & inp_batch) {
 
                     GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits
 
-                    for (int32_t i = 0; i < n_tokens; i++) {
+                    // TODO: fix indexing [UBATCH_IDX]
+                    for (uint32_t i = 0; i < n_tokens; i++) {
                         const llama_seq_id seq_id = ubatch.seq_id[i][0];
                         if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
                             continue;
@@ -846,6 +837,7 @@ int llama_context::encode(llama_batch & inp_batch) {
                     auto & embd_seq_out = embd_seq;
                     const uint32_t n_cls_out = hparams.n_cls_out;
 
+                    // TODO: fix indexing [UBATCH_IDX]
                     for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
                         const llama_seq_id seq_id = ubatch.seq_id[s][0];
                         if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
@@ -878,13 +870,11 @@ int llama_context::encode(llama_batch & inp_batch) {
         memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd));
 
         // remember the sequence ids used during the encoding - needed for cross attention later
-        // TODO: the seuqence indexing here is likely not correct in the general case
-        //       probably works only for split_simple
         cross.seq_ids_enc.resize(n_tokens);
-        for (int32_t i = 0; i < n_tokens; i++) {
+        for (uint32_t i = 0; i < n_tokens; i++) {
             cross.seq_ids_enc[i].clear();
-            for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
-                llama_seq_id seq_id = ubatch.seq_id[i][s];
+            for (int s = 0; s < batch.n_seq_id[i]; s++) {
+                llama_seq_id seq_id = batch.seq_id[i][s];
                 cross.seq_ids_enc[i].insert(seq_id);
             }
         }
@@ -893,68 +883,43 @@ int llama_context::encode(llama_batch & inp_batch) {
     return 0;
 }
 
-int llama_context::decode(llama_batch & inp_batch) {
+int llama_context::decode(const llama_batch & batch_inp) {
     if (!memory) {
         LLAMA_LOG_DEBUG("%s: cannot decode batches with this context (calling encode() instead)\n", __func__);
-        return encode(inp_batch);
+        return encode(batch_inp);
     }
 
-    if (inp_batch.n_tokens == 0) {
+    if (batch_inp.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
-    if (!inp_batch.pos) {
-        if (inp_batch.seq_id) {
-            LLAMA_LOG_ERROR("%s: pos == NULL, but seq_id != NULL\n", __func__);
-            return -1;
-        }
+    if (!batch_allocr->init(batch_inp, model.vocab, memory.get())) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+        return -1;
     }
 
-    // temporary allocate memory for the input batch if needed
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : memory->seq_pos_max(0) + 1);
-
-    const llama_batch & batch = batch_allocr.batch;
+    const llama_batch & batch = batch_allocr->get_batch();
 
     const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
 
     const int32_t n_vocab = vocab.n_tokens();
+    const int64_t n_embd  = hparams.n_embd;
 
-    const int64_t n_tokens_all = batch.n_tokens;
-    const int64_t n_embd       = hparams.n_embd;
+    const uint32_t n_tokens_all = batch.n_tokens;
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    // TODO: move the validation to the llama_batch_allocr
-    if (batch.token) {
-        for (int64_t i = 0; i < n_tokens_all; ++i) {
-            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
-                LLAMA_LOG_ERROR("%s: invalid token[%" PRId64 "] = %d\n", __func__, i, batch.token[i]);
-                return -1;
-            }
-
-            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
-                LLAMA_LOG_ERROR("%s: invalid seq_id[%" PRId64 "] = %d >= %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
-                return -1;
-            }
-        }
-    }
-
     // this indicates we are doing pooled embedding
     const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
 
-    int64_t n_outputs_all = 0;
-
-    // count outputs
-    for (uint32_t i = 0; i < n_tokens_all; ++i) {
-        n_outputs_all += batch.logits[i] != 0;
-    }
+    const uint32_t n_outputs_all = batch_allocr->get_n_outputs();
 
     if (embd_pooled) {
         // require that all tokens are output
         if (n_outputs_all != n_tokens_all) {
-            LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %" PRId64 ", n_tokens_all = %" PRId64 ")\n",
+            LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %d, n_tokens_all = %d)\n",
                     __func__, n_outputs_all, n_tokens_all);
             return -1;
         }
@@ -1024,7 +989,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 
     // reserve output buffer
     if (output_reserve(n_outputs_all) < n_outputs_all) {
-        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
+        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
         return -2;
     };
 
@@ -1058,18 +1023,19 @@ int llama_context::decode(llama_batch & inp_batch) {
 
         if (!res) {
             // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
-            llama_pos pos_min[LLAMA_MAX_PARALLEL_SEQUENCES];
-            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+            llama_pos pos_min[LLAMA_MAX_SEQ];
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                 pos_min[s] = std::numeric_limits<llama_pos>::max();
             }
 
+            // TODO: fix sequence indexing
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 const auto & seq_id = ubatch.seq_id[i][0];
 
                 pos_min[seq_id] = std::min(pos_min[seq_id], ubatch.pos[i]);
             }
 
-            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                 if (pos_min[s] == std::numeric_limits<llama_pos>::max()) {
                     continue;
                 }
@@ -1176,14 +1142,14 @@ int llama_context::decode(llama_batch & inp_batch) {
     n_outputs = n_outputs_all;
 
     // set output mappings
-    {
+    if (n_outputs > 0) {
         bool sorted_output = true;
 
         auto & out_ids = mstate->out_ids();
 
-        GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all);
+        GGML_ASSERT(out_ids.size() == (size_t) n_outputs);
 
-        for (int64_t i = 0; i < n_outputs_all; ++i) {
+        for (int64_t i = 0; i < n_outputs; ++i) {
             int64_t out_id = out_ids[i];
             output_ids[out_id] = i;
             if (out_id != i) {
@@ -1195,20 +1161,22 @@ int llama_context::decode(llama_batch & inp_batch) {
         // note: this is mostly relevant for recurrent models atm
         if (!sorted_output) {
             const uint32_t n_vocab = model.vocab.n_tokens();
-            const uint32_t n_embd  = model.hparams.n_embd;
+            const uint64_t n_embd  = model.hparams.n_embd;
 
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
             // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort)
-            for (int32_t i = 0; i < n_outputs - 1; ++i) {
-                int32_t j_min = i;
-                for (int32_t j = i + 1; j < n_outputs; ++j) {
+            for (uint32_t i = 0; i < n_outputs - 1; ++i) {
+                uint32_t j_min = i;
+                for (uint32_t j = i + 1; j < n_outputs; ++j) {
                     if (out_ids[j] < out_ids[j_min]) {
                         j_min = j;
                     }
                 }
-                if (j_min == i) { continue; }
+                if (j_min == i) {
+                    continue;
+                }
                 std::swap(out_ids[i], out_ids[j_min]);
                 if (logits_size > 0) {
                     for (uint32_t k = 0; k < n_vocab; k++) {
@@ -1221,8 +1189,10 @@ int llama_context::decode(llama_batch & inp_batch) {
                     }
                 }
             }
+
             std::fill(output_ids.begin(), output_ids.end(), -1);
-            for (int32_t i = 0; i < n_outputs; ++i) {
+
+            for (uint32_t i = 0; i < n_outputs; ++i) {
                 output_ids[out_ids[i]] = i;
             }
         }
@@ -1242,7 +1212,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 // output
 //
 
-int32_t llama_context::output_reserve(int32_t n_outputs) {
+uint32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1308,8 +1278,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     // set all ids as invalid (negative)
     std::fill(output_ids.begin(), output_ids.end(), -1);
 
-    this->n_outputs     = 0;
-    this->n_outputs_max = n_outputs_max;
+    this->n_outputs = 0;
 
     return n_outputs_max;
 }
@@ -1800,14 +1769,12 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
 
         std::vector<int32_t> w_output_pos;
 
-        GGML_ASSERT(n_outputs <= n_outputs_max);
-
         w_output_pos.resize(n_outputs);
 
         // build a more compact representation of the output ids
         for (size_t i = 0; i < n_batch(); ++i) {
             // map an output id to a position in the batch
-            int32_t pos = output_ids[i];
+            int64_t pos = output_ids[i];
             if (pos >= 0) {
                 GGML_ASSERT(pos < n_outputs);
                 w_output_pos[pos] = i;
@@ -2082,7 +2049,7 @@ void llama_context::opt_epoch_iter(
 
         embd_seq.clear();
 
-        int64_t n_outputs_all = n_tokens_all;
+        uint32_t n_outputs_all = n_tokens_all;
 
         auto mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled);
         if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
@@ -2092,7 +2059,7 @@ void llama_context::opt_epoch_iter(
 
         // reserve output buffer
         if (output_reserve(n_outputs_all) < n_outputs_all) {
-            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
+            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
 

src/llama-context.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #include "llama.h"
-#include "llama-batch.h"
 #include "llama-cparams.h"
 #include "llama-graph.h"
 #include "llama-adapter.h"
@@ -13,6 +12,7 @@
 #include <vector>
 
 struct llama_model;
+class llama_batch_allocr;
 
 class llama_io_read_i;
 class llama_io_write_i;
@@ -102,8 +102,8 @@ struct llama_context {
             llama_memory_state_i * mstate,
                      ggml_status & ret);
 
-    int encode(llama_batch & inp_batch);
-    int decode(llama_batch & inp_batch);
+    int encode(const llama_batch & batch_inp);
+    int decode(const llama_batch & batch_inp);
 
     //
     // state save/load
@@ -181,7 +181,7 @@ private:
 
     // Make sure enough space is available for outputs.
     // Returns max number of outputs for which space was reserved.
-    int32_t output_reserve(int32_t n_outputs);
+    uint32_t output_reserve(int32_t n_outputs);
 
     //
     // graph
@@ -246,8 +246,10 @@ private:
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
-    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
+    // reuse the batch_allocr to avoid unnecessary memory allocations
+    std::unique_ptr<llama_batch_allocr> batch_allocr;
+
+    uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch
 
     std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 

src/llama-cparams.cpp

@@ -1,5 +1,5 @@
 #include "llama-cparams.h"
 
 size_t llama_max_parallel_sequences(void) {
-    return LLAMA_MAX_PARALLEL_SEQUENCES;
+    return LLAMA_MAX_SEQ;
 }

src/llama-cparams.h

@@ -4,7 +4,7 @@
 
 #include <cstdint>
 
-#define LLAMA_MAX_PARALLEL_SEQUENCES 64
+#define LLAMA_MAX_SEQ 64
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference

src/llama-graph.cpp

@@ -139,6 +139,7 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
 
         std::vector<uint64_t> sum(n_tokens, 0);
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -156,6 +157,7 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
             }
         }
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -180,6 +182,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
         uint32_t * data = (uint32_t *) cls->data;
         memset(cls->data, 0, n_tokens * ggml_element_size(cls));
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -210,6 +213,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
         std::vector<int> last_pos(n_tokens, -1);
         std::vector<int> last_row(n_tokens, -1);
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -283,6 +287,7 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
+                                // TODO: fix indexing [UBATCH_IDX]
                                 for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
                                     if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) {
                                         if (hparams.use_alibi) {
@@ -322,6 +327,7 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
+                                // TODO: fix indexing [UBATCH_IDX]
                                 for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
                                     if (ubatch->seq_id[s0][s] == seq_id) {
                                         if (hparams.use_alibi) {
@@ -377,6 +383,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
             for (int j = 0; j < n_tokens; ++j) {
                 for (int i = 0; i < n_enc; ++i) {
                     float f = -INFINITY;
+                    // TODO: fix indexing [UBATCH_IDX]
                     for (int s = 0; s < ubatch->n_seq_id[j]; ++s) {
                         const llama_seq_id seq_id = ubatch->seq_id[j][s];
                         if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) {
@@ -1556,23 +1563,30 @@ void llm_graph_context::build_pooling(
                 ggml_tensor * inp_cls = build_inp_cls();
                 inp = ggml_get_rows(ctx0, inp, inp_cls);
 
-                if (cls != nullptr && cls_b != nullptr) {
+                if (cls) {
                     // classification head
                     // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
+                    cur = ggml_mul_mat(ctx0, cls, inp);
+                    if (cls_b) {
+                        cur = ggml_add(ctx0, cur, cls_b);
+                    }
                     cur = ggml_tanh(ctx0, cur);
 
                     // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
                     // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
                     if (cls_out) {
-                        GGML_ASSERT(cls_out_b != nullptr);
-                        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
+                        cur = ggml_mul_mat(ctx0, cls_out, cur);
+                        if (cls_out_b) {
+                            cur = ggml_add(ctx0, cur, cls_out_b);
+                        }
                     }
                 } else if (cls_out) {
                     // Single layer classification head (direct projection)
                     // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
-                    GGML_ASSERT(cls_out_b != nullptr);
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
+                    cur = ggml_mul_mat(ctx0, cls_out, inp);
+                    if (cls_out_b) {
+                        cur = ggml_add(ctx0, cur, cls_out_b);
+                    }
                 } else {
                     GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
                 }

src/llama-graph.h

@@ -378,7 +378,7 @@ struct llm_graph_params {
     const llama_memory_state_i * mstate;
     const llama_cross          * cross;
 
-    int32_t n_outputs;
+    uint32_t n_outputs;
 
     const llm_graph_cb & cb;
 };
@@ -412,8 +412,8 @@ struct llm_graph_context {
     const float norm_eps;
     const float norm_rms_eps;
 
-    const int32_t n_tokens;
-    const int32_t n_outputs;
+    const int64_t n_tokens;
+    const int64_t n_outputs;
     const int32_t n_ctx_orig; // yarn
 
     const enum llama_pooling_type pooling_type;

src/llama-kv-cache-unified.cpp

@@ -572,7 +572,7 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
             LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
         }
 
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (cells.seq_pos_min(s) < 0) {
                 continue;
             }
@@ -652,8 +652,8 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch
 
     // keep track of the max sequence position that we would overwrite with this ubatch
     // for non-SWA cache, this would be always empty
-    llama_seq_id seq_pos_max_rm[LLAMA_MAX_PARALLEL_SEQUENCES];
-    for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+    llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ];
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
         seq_pos_max_rm[s] = -1;
     }
 
@@ -674,6 +674,7 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch
 
             cells.pos_set(head_cur + idx, ubatch.pos[idx]);
 
+            // TODO: fix indexing [UBATCH_IDX]
             for (int32_t i = 0; i < ubatch.n_seq_id[s]; i++) {
                 cells.seq_add(head_cur + idx, ubatch.seq_id[s][i]);
             }
@@ -683,7 +684,7 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch
     // note: we want to preserve the invariant that all positions between [pos_min, pos_max] for each sequence
     //       will be present in the cache. so we have to purge any position which is less than those we would overwrite
     //       ref: https://github.com/ggml-org/llama.cpp/pull/13746#issuecomment-2916057092
-    for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
         if (seq_pos_max_rm[s] == -1) {
             continue;
         }
@@ -1738,7 +1739,7 @@ llama_kv_cache_unified_state::llama_kv_cache_unified_state(
         llama_context * lctx,
         bool do_shift,
         defrag_info dinfo) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), lctx(lctx), do_shift(do_shift), dinfo(std::move(dinfo)) {
-    if (!do_shift && dinfo.empty()) {
+    if (!do_shift && this->dinfo.empty()) {
         status = LLAMA_MEMORY_STATUS_NO_UPDATE;
     }
 }

src/llama-kv-cells.h

@@ -23,7 +23,7 @@ public:
 
         used.clear();
 
-        for (uint32_t s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
             seq_pos[s].clear();
         }
     }
@@ -240,7 +240,7 @@ public:
     llama_seq_id seq_get(uint32_t i) const {
         assert(seq[i].count() == 1);
 
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 return s;
             }
@@ -253,7 +253,7 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_min(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -266,7 +266,7 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_max(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -384,20 +384,20 @@ private:
     //
     std::vector<llama_pos> shift;
 
-    using bits_t = std::bitset<LLAMA_MAX_PARALLEL_SEQUENCES>;
+    using bits_t = std::bitset<LLAMA_MAX_SEQ>;
 
     // the bitset seq[i] tells us which sequences are currently occupying the i-th cell
     std::vector<bits_t> seq;
 
     // the set seq_pos[s] tells us which positions are currently present for sequence s
     // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache
-    std::set<llama_pos> seq_pos[LLAMA_MAX_PARALLEL_SEQUENCES];
+    std::set<llama_pos> seq_pos[LLAMA_MAX_SEQ];
 
     // helper functions for updating `seq_pos`, once cell at a time:
 
     // remove cell i
     void seq_pos_rm(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].erase(pos[i]);
             }
@@ -406,7 +406,7 @@ private:
 
     // add cell i
     void seq_pos_add(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].insert(pos[i]);
             }

src/llama-model.cpp

@@ -80,6 +80,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_40B:           return "40B";
         case LLM_TYPE_65B:           return "65B";
         case LLM_TYPE_70B:           return "70B";
+        case LLM_TYPE_142B:          return "142B";
         case LLM_TYPE_236B:          return "236B";
         case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_314B:          return "314B";
@@ -1444,6 +1445,20 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_DOTS1:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+                switch (hparams.n_layer) {
+                    case 62: type = LLM_TYPE_142B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -4123,6 +4138,58 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
                     }
                 } break;
+            case LLM_ARCH_DOTS1:
+                {
+                    const int64_t n_ff_exp        = hparams.n_ff_exp;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (i < (int) hparams.n_layer_dense_lead) {
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        } else {
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+                            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+                            if (n_expert == 0) {
+                                throw std::runtime_error("n_expert must be > 0");
+                            }
+                            if (n_expert_used == 0) {
+                                throw std::runtime_error("n_expert_used must be > 0");
+                            }
+
+                            // MoE branch
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                            // Shared expert branch
+                            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+                            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                        }
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -13194,6 +13261,156 @@ struct llm_build_bailingmoe : public llm_graph_context {
     }
 };
 
+struct llm_build_dots1 : public llm_graph_context {
+    llm_build_dots1(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            if ((uint32_t) il < hparams.n_layer_dense_lead) {
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                ggml_tensor * moe_out =
+                    build_moe_ffn(cur,
+                            model.layers[il].ffn_gate_inp,
+                            model.layers[il].ffn_up_exps,
+                            model.layers[il].ffn_gate_exps,
+                            model.layers[il].ffn_down_exps,
+                            model.layers[il].ffn_exp_probs_b,
+                            n_expert, n_expert_used,
+                            LLM_FFN_SILU, hparams.expert_weights_norm,
+                            true, hparams.expert_weights_scale,
+                            (llama_expert_gating_func_type) hparams.expert_gating_func,
+                            il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                {
+                    ggml_tensor * ffn_shexp = build_ffn(cur,
+                            model.layers[il].ffn_up_shexp,   NULL, NULL,
+                            model.layers[il].ffn_gate_shexp, NULL, NULL,
+                            model.layers[il].ffn_down_shexp, NULL, NULL,
+                            NULL,
+                            LLM_FFN_SILU, LLM_FFN_PAR, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                }
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
@@ -13532,6 +13749,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_bailingmoe>(*this, params, gf);
             } break;
+        case LLM_ARCH_DOTS1:
+            {
+                llm = std::make_unique<llm_build_dots1>(*this, params, gf);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -13714,6 +13935,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_NEMOTRON:
         case LLM_ARCH_EXAONE:
         case LLM_ARCH_MINICPM3:
+        case LLM_ARCH_DOTS1:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:

src/llama-model.h

@@ -73,6 +73,7 @@ enum llm_type {
     LLM_TYPE_40B,
     LLM_TYPE_65B,
     LLM_TYPE_70B,
+    LLM_TYPE_142B,
     LLM_TYPE_236B,
     LLM_TYPE_290B,
     LLM_TYPE_314B,

src/llama-quant.cpp

@@ -585,7 +585,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
-                gguf_set_val_i32(ctx_out.get(), o.key, o.val_i64);
+                // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {

src/llama-vocab.cpp

@@ -9,16 +9,16 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cctype>
 #include <cfloat>
-#include <climits>
 #include <cstdarg>
 #include <cstring>
 #include <forward_list>
+#include <limits>
 #include <map>
 #include <queue>
 #include <set>
 #include <unordered_map>
-#include <cctype>
 
 //
 // helpers
@@ -2572,6 +2572,10 @@ int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t
     // copy piece chars to output text buffer
     // skip up to 'lstrip' leading spaces before copying
     auto _try_copy = [=] (const char * token, size_t size) -> int32_t {
+        if (size >= static_cast<size_t>(std::numeric_limits<int32_t>::max())) {
+            GGML_ABORT("invalid token size: %zu exceeds int32_t limit", size);
+        }
+
         for (int32_t i = 0; i < lstrip && size && *token == ' '; ++i) {
             token++;
             size--;

tools/server/server.cpp

@@ -3217,7 +3217,7 @@ struct server_context {
                                 }
 
                                 const auto n_swa = llama_model_n_swa(model);
-                                if (pos_min > slot.n_past - n_swa) {
+                                if (pos_min > std::max(0, slot.n_past - n_swa)) {
                                     SLT_WRN(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.cache_tokens.size(), slot.id, pos_min, n_swa);
                                     SLT_WRN(slot, "forcing full prompt re-processing due to lack of cache data (likely due to SWA, see %s)\n",
                                             "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
@@ -4878,7 +4878,9 @@ int main(int argc, char ** argv) {
     };
 
     bool was_bound = false;
+    bool is_sock = false;
     if (string_ends_with(std::string(params.hostname), ".sock")) {
+        is_sock = true;
         LOG_INF("%s: setting address family to AF_UNIX\n", __func__);
         svr->set_address_family(AF_UNIX);
         // bind_to_port requires a second arg, any value other than 0 should
@@ -4956,7 +4958,9 @@ int main(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-    LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
+    LOG_INF("%s: server is listening on %s - starting the main loop\n", __func__,
+            is_sock ? string_format("unix://%s", params.hostname.c_str()).c_str() :
+                      string_format("http://%s:%d", params.hostname.c_str(), params.port).c_str());
 
     // this call blocks the main thread until queue_tasks.terminate() is called
     ctx_server.queue_tasks.start_loop();